Introduced AI Hospital multi-agent framework; models achieve <50% of GPT-4 one-step accuracy; highlights gaps in multi-turn clinical reasoning.

Zhihao Fan, Jialong Tang, Wei Chen et al.

DoRA enhances LoRA's learning capacity via weight decomposition, outperforming LoRA on multiple tasks.

Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin et al.

SLGM model enhances structure prediction via loss calibration and formatted decoding, significantly boosting <1B parameter model performance.

Minho Lee, Junghyun Min, Yerang Kim et al.

InternLM-Math unifies chain-of-thought, reward modeling, formal proof, and code interpretation, achieving SOTA on multiple benchmarks.

Huaiyuan Ying, Shuo Zhang, Linyang Li et al.

Introduces WebLINX, a multi-turn dialogue benchmark for real-world website navigation, using a retrieval-based model to select relevant HTML elements, achieving 23.9 score on OOD tests.

Xing Han Lù, Zdeněk Kasner, Siva Reddy

The paper introduces the multilingual E5 (mE5) text embedding models trained on over 1 billion multilingual text pairs using contrastive learning, combined with supervised fine-tuning and instruction tuning, achieving state-of-the-art performance in multilingual retrieval benchmarks.

Liang Wang, Nan Yang, Xiaolong Huang et al.

DeepSeekMath 7B通过120B数学相关tokens和GRPO算法，显著提升开源模型数学推理能力，接近GPT-4表现。

Zhihong Shao, Peiyi Wang, Qihao Zhu et al.

Study shows LLMs can perform table-based fact verification with prompt engineering and instruction tuning.

Hanwen Zhang, Qingyi Si, Peng Fu et al.

PiCO uses consistency optimization for unsupervised peer review among LLMs, achieving rankings closer to human preferences.

Kun-Peng Ning, Shuo Yang, Yu-Yang Liu et al.

Using Dutch image descriptions and eye-tracking data, this study quantifies human visuo-linguistic signal variation and evaluates pretrained models' ability to capture it.

Ece Takmaz, Sandro Pezzelle, Raquel Fernández

LitLLM toolkit uses Retrieval Augmented Generation to significantly reduce literature review time.

Shubham Agarwal, Gaurav Sahu, Abhay Puri et al.

Proposes CodeAct, a framework enabling LLMs to generate executable Python code for actions, significantly improving success rates (up to 20%) in complex multi-tool tasks via integrated Python interpreter and multi-turn interactions.

Xingyao Wang, Yangyi Chen, Lifan Yuan et al.

Superfiltering uses small models like GPT-2 to efficiently select high-quality instruction data for large models, boosting performance and reducing costs.

Ming Li, Yong Zhang, Shwai He et al.

Combining Chain-of-Thought prompting with external symbolic tools, the model achieved 85% accuracy on MathQA, significantly advancing AI mathematical reasoning.

Janice Ahn, Rishu Verma, Renze Lou et al.

RAPTOR constructs recursive abstracted trees for long document retrieval, achieving 20%+ accuracy gains on QA benchmarks.

Parth Sarthi, Salman Abdullah, Aditi Tuli et al.

This study benchmarks GPT-4V on Earth observation tasks, revealing strong scene understanding but poor spatial reasoning for localization and counting.

Chenhui Zhang, Sherrie Wang

Mobile-Agent uses visual perception with DINO, OCR, and GPT-4V to autonomously operate mobile apps, achieving 91% success in Mobile-Eval.

Junyang Wang, Haiyang Xu, Jiabo Ye et al.

Routoo uses a lightweight predictor and cost-aware routing to reduce inference costs by 33% while maintaining performance.

Alireza Mohammadshahi, Arshad Rafiq Shaikh, Majid Yazdani

WebVoyager achieves a 59.1% task success rate using large multimodal models.

Hongliang He, Wenlin Yao, Kaixin Ma et al.

SEER enhances structured reasoning via reinforcement learning, achieving a 6.9% improvement on EntailmentBank.

Guoxin Chen, Kexin Tang, Chao Yang et al.